DUCATI: A Dual-Cache Training System for Graph Neural Networks on Giant Graphs with the GPU | Synapse